Thai Rubber Leaf Disease Classification Using Deep Learning Techniques

This research investigated the utilization of deep learning methodologies for the precise classification of Thai rubber leaf diseases using leaf images. The study employed a diverse set of convolutional neural network (CNN) models, including VGG16, ResNet50, EfficientNet, and YOLOv8, to accomplish this task. Data augmentation techniques were employed to address overfitting, thereby enhancing the robustness of the models. The dataset, meticulously curated and labeled by domain experts, comprised 734 high-quality rubber leaf images categorized into three distinct classes: new diseases, powdery mildew disease, and healthy leaves. The findings revealed that YOLOv8n-cls emerged as the optimal choice, demonstrating exceptional efficiency with a diminutive model size (5.5 MB) and a parsimonious parameter count (1.44 million). This model exhibited remarkable accuracy (98.68%) and an impressive F1-score of 0.98, establishing its proficiency in delivering precise and reliable predictions. Furthermore, a comparative analysis of YOLOv8 pretrained classify models underscored YoloV8s as the most favorable option, characterized by the lowest validation loss. These results highlighted the potential of deep learning methodologies in the field of agricultural disease detection, offering a cost-effective and efficient solution for rubber plantation owners and farmers. This research contributed valuable insights into the application of advanced technology to address critical challenges in the agricultural sector, particularly in the context of plant disease detection and classification.


INTRODUCTION
The rubber industry has experienced profound changes over its historical progression, with these alterations predominantly shaped by the rubber boom of the early 20th century, largely driven by trading enterprises [1].Natural rubber (NR) holds paramount significance due to its extensive utilization across diverse applications and products, conferring it with profound international import.Noteworthy scholars such as Fong, Khin, and Lim have conducted comprehensive investigations into NR production, consumption, and pricing models within specific ASEAN countries and the global market [2].Furthermore, Prasada and Dhamira have illuminated the escalating global demand for natural rubber, alongside the competitive panorama among exporting nations like Thailand, Indonesia, and Malaysia [3].Notably, Perdana underscores the pivotal import of nurturing downstream rubber industry growth to stabilize rubber prices and enhance domestic consumption, particularly within the context of Indonesia [4].Moreover, Gayan accentuates the socioeconomic advantages intrinsic to rubber plantations, encompassing facets such as employment generation and the sequestration of carbon emissions [5].
Rubber plantations hold a significant role in the agricultural domain due to their contribution to the global latex demand and their role as a livelihood source for a substantial population [6,7].Nonetheless, the proliferation of rubber plantations has given rise to unfavorable consequences concerning biodiversity and ecosystem functionalities [8].Empirical inquiries underscore that rubber plantations generally exhibit diminished ecosystem functionalities in comparison to forests, characterized by noteworthy reductions in both aboveground and belowground biomass, alongside a decline in plant diversity [9].Leaf afflictions bear noteworthy implications for rubber output and quality.In China, examined factors affecting rubber latex yield.The region's climate and powdery mildew reduce yield by 20% [10].Analogously, within Cameroon, leaf diseases instigated by Fusarium oxysporum and Pestalotiopsis microspora have contributed to a reduction in latex yield [11]. Noteworthy among these is Corynespora leaf fall, attributed to Corynespora cassiicola, which emerges as a particularly pernicious leaf ailment affecting Hevea rubber, resulting in a substantial 45% reduction in yield [12]. Leaf diseases, as observed across diverse plants including rubber, lead to significant reductions in production and economic returns, accompanied by a diminution in both the quality and quantity of agricultural yield [13]. Proficient management of foliar diseases, encompassing conditions like crown rust and leaf spot, plays a pivotal role in curtailing yield and grain quality reductions within oat crops [14].
In the context of Thailand, the rubber sector assumes significant economic importance for the nation and its agricultural community.As the leading global exporter of natural rubber, contributing onethird of the world's production, the industry confronts challenges stemming from leaf diseases, which pose threats to its sustainable growth.To address this issue, experts have identified the adoption of rubber crop diversity systems as a potential solution, providing a means for farmers to mitigate negative impacts and secure their income [15].Several factors influence the adoption of rubber crop diversity, encompassing management skills, access to information, market opportunities, rainfall patterns, water availability, household labor, land rights, land degradation, use of organic fertilizers, crop sales channels, and irrigation access [16].In order to enhance the competitiveness of Thailand's rubber industry while concomitantly mitigating its ecological impact, policymakers are obligated to concentrate their efforts on specific areas.Moreover, it is imperative to recognize that the proficient management of rubber leaf disease presents a substantial challenge within this undertaking.Primarily, raising awareness and knowledge of sustainable practices among rubber farmers is paramount.Secondly, efforts to minimize firewood usage and promote energy efficiency will contribute to a more sustainable industry.Lastly, providing support for infrastructure development can further enhance the industry's overall performance and sustainability [17,18].
Deep learning methodologies encompass a diverse spectrum of promising applications within the agricultural sector.They exhibit effectiveness in the analysis of hyperspectral images, facilitating tasks such as ripeness prediction, component identification, classification, and the detection of plant diseases [19].Additionally, deep learning strategies offer valuable solutions to address complexities associated with plant identification, classification, and segmentation.This is achieved by harnessing self-supervised contrastive learning on agricultural images [20].Moreover, deep learning algorithms assume an indispensable role as tools for decision support systems in agriculture, facilitating critical functions such as yield estimation, crop disease discernment, weed identification, and irrigation scheduling [21].Furthermore, deep learning has demonstrated its efficacy in the arena of horticultural research, catering to a multitude of objectives encompassing variety recognition, yield estimation, quality assessment, stress phenotyping detection, and growth monitoring [22].The demonstrated efficacy in these applications underscores the considerable potential of deep learning methodologies in propelling agricultural practices and providing vital assistance to farmers in their pursuits to monitor crops and enhance yields [7].In light of the substantial impact of plant diseases on agricultural production and food quality, the need for efficient plant disease detection and classification is paramount [23].The limitations inherent in conventional manual diagnosis techniques, owing to accuracy concerns and the dependency on human resources, can be circumvented through the proposition of automated strategies utilizing deep learning algorithms [24,25].
The primary research objective of this study is to develop a deep learning-based approach for accurate classification of Thai rubber leaf disease using leaf images.The proposed approach aims to enhance the efficiency and accuracy of disease classification, offering a cost-effective solution for rubber plantation owners and farmers.Employing a convolutional neural network (CNN) architecture trained on a large dataset of rubber leaf images, the model effectively detects the presence of the disease and classifies the severity level, vital for formulating appropriate management strategies.The rest of the segments are organized as follows: the second section discusses the related work.The data collection and preprocessing for leaf disease classification is described in section 3. The proposed method was tested for efficiency and performance in section 4, and the result was presented in tables and graphs.The discussion and conclusions are summarized in the final section.

RELATED WORK
The effective management of rubber leaf disease poses a significant challenge to the rubber plantation industry, necessitating the adoption of early detection and accurate classification methods.Previous studies have explored spectral imaging and machine learning techniques for disease detection, although these methods often demand specialized equipment and extensive data.In recent years, image classification techniques in plant pathology have undergone extensive investigation, with numerous methodologies proposed to automate disease diagnosis and enhance classification accuracy.Convolutional neural networks (CNNs) have emerged as a prominent machine learning approach, gaining widespread adoption for their efficacy in plant disease detection [26,27].These techniques have demonstrated successful application to diverse image types, encompassing those captured through visible light cameras and sensors capturing non-visible wavelengths [28].Researchers have delved into supervised, unsupervised, and semi-supervised classification methodologies, employing various techniques, including transfer learning, support vector machines, k-nearest neighbor, and random forest algorithms, in the context of image classification for plant pathology [29].The integration of computer vision and neural networks, particularly emphasizing the utilization of CNNs, presents substantial potential for advancing disease detection in plants [30].
In the specific domain of leaf disease detection and classification, diverse techniques, such as spectral imaging, machine learning algorithms, and deep learning models, have been explored.While these methods have demonstrated promising results, they are not without limitations in terms of accuracy, speed, and cost-effectiveness.For instance, Chenghai Yin et al. proposed a deep learning network, named DISE-Net, for the classification of maize small leaf spot disease.They achieved superior accuracy (97.12%) in comparison .These studies underscore the potential of deep learning in disease detection and highlight its capacity to assist farmers in timely disease treatment, ultimately resulting in enhancements in both crop quality and quantity [32].Furthermore, Kantip Kiratiratanapruk et al. demonstrated the potential of machine vision technology and CNNs in detecting and identifying six major rice diseases, showcasing the impact of such techniques on enhancing agricultural production and sustaining global food security [33].
In the context of rubber leaf disease, recent studies have explored deep learning-based approaches.For instance, Tiwei Zeng et al. introduced the GMA-Net model for rubber leaf disease recognition, achieving remarkable recognition accuracies of 98.06% and 99.43% [34].Additionally, other researchers have evaluated the performance of various CNN models, such as VGG16, Faster R-CNN, ResNet50, InceptionV3, YOLOv3, and EfficientNet, in diagnosing and identifying rubber leaf diseases in Thailand [33].The results underscore the feasibility and potential of CNNs in detecting and classifying rubber leaf diseases at various stages and severities.Furthermore, researchers are planning to develop a Rubber Leaf Disease diagnostic system and a mobile application to provide an alternative tool to traditional methods reliant on plant disease specialists.In conclusion, numerous studies have demonstrated the effectiveness of deep learning techniques, particularly CNNs, in the early detection and accurate classification of plant diseases, including rubber leaf disease.The advancements in deep learning hold promise for transforming disease management practices in the agricultural sector, enhancing crop yield, and contributing to sustainable agricultural practices.As the field continues to progress, future research endeavors aim to optimize models, explore mobilebased applications, and expand datasets to encompass diverse types of rubber leaf diseases and other crop ailments.

MATERIALS AND METHODS 3.1 Materials
3.1.1Data acquisition.The Rubber Leaf Disease Dataset was meticulously assembled, comprising a total of 734 rubber leaf samples meticulously collected from diverse locations, namely Rubber Authority of Thailand, Rubber Research Center, and Faculty of Natural Resources, Prince of Songkla University.Each sample was rigorously assessed by domain experts, who possessed advanced knowledge in rubber leaf diseases, enabling precise labeling and classification.The categorization and labeling of various rubber leaf diseases were exclusively based on distinct external visual symptoms.Subsequently, image data were systematically collected within an authentic environment.A substantial portion of these sample images were obtained through the utilization of diverse smartphone models with varying operating systems.
We concentrated on gathering image samples representing only three distinct categories of rubber leaves: those afflicted by new diseases, those suffering from powdery mildew disease, and those considered as healthy.In Table 1 and Table 2, we present illustrative examples of typical symptoms associated with these specific rubber leaf diseases.

Data set preparation.
Prior to image data augmentation, a series of essential image preprocessing steps were meticulously executed on the original images.These preprocessing procedures Subsequently, these lesions progress to encompass the entire leaf.As the lesions mature, the powdery mildew spots transform into circular, ringworm-like formations with a white appearance.Concurrently, the leaf's surface undergoes desiccation and takes on a yellow hue, culminating in its eventual detachment.New Disease In the early stages of symptom development, a discernible bruised lesion emerges beneath the leaf, accompanied by the appearance of a circular yellowing on the leaf's upper surface within the same vicinity.Subsequently, this area undergoes expansion, with the wound edges darkening and transitioning into a brown, desiccated tissue that ultimately fades into pale white lines.The wound typically maintains a roughly circular shape, without a surrounding yellow halo; multiple points of affliction may converge, forming a larger wound.In cases of severe symptomatology, the leaves will exhibit yellowing and eventual abscission.
encompassed image scaling, clipping, and the removal of extraneous image backgrounds.In this study, a uniform resizing of sample images to dimensions of 224 x 224 pixels was exclusively adopted for input into the image analysis pipeline.This resizing not only reduced computational expenses but also significantly improved the overall efficiency of image processing.
Our dataset comprises a total of 734 images, categorized as follows: 205 instances of powdery mildew disease, 259 instances of new disease, and 270 instances of healthy leaves.During the dataset preparation for the training phase, we divided our dataset into distinct subsets: 70% for the training set, 20% for the validation set, and 10% for the test set.An analysis of sample distribution across these categories highlights the dataset's substantial size.Consequently, we employed an image data augmentation strategy to expand the training dataset by introducing modified replicas of existing data instances, increasing the size of the training set to 1,536 images.This augmentation approach serves a dual purpose: not only does it increase the dataset's size, but it also acts as a regularization technique, helping to mitigate potential overfitting.In this research paper, we set the batch size to 32.To enhance diversity within our image dataset, we employed various image augmentation techniques, including brightness adjustment, rotation, scaling, horizontal flipping, and other methods.

Methods
In our methods, we aimed to locate and classify an object of interest in images using four common classification models: VGG16, RestNet50, EfficientNet, and YOLOv8.As deep learning network training requires a large dataset and is time-consuming, we used a transfer learning technique with pre-trained models from ImageNet to initialize the weights.We then retrained the pre-trained model with our target dataset to update the weights based on the object types we wanted to classify and adjusted the learning parameters to suit our purposes (Figure 1).
Another issue we considered during Convolutional Neural Network (CNN) training was overfitting, which occurs when the model learns well on the training dataset but fails to generalize to new data not included in the training set.To address this problem, we applied an augmentation technique to increase the amount of training data, using various transformations such as random flipping, scaling, rotation, translation, brightness, contrast, and hue.This technique helped the model learn patterns in sample data during the training process and reduced the overfitting problem [33].

VGG16.
The VGG16 is a CNN model with 16 layers proposed by K. Simonyan and A. Zisserman from Oxford University [35].It achieved a test accuracy of 92.7% in the ImageNet dataset and is one of the top models in ILSVRC-2014 competition.VGG16 uses smaller 3x3 filters in place of larger filters used in AlexNet and was trained for several weeks using NVIDIA Titan Black GPUs.

ResNet50.
The ResNet-50 is a pre-trained Convolutional Neural Network (CNN) for image classification composed of 50 layers, trained on a dataset of one million images from the ImageNet database.With over 23 million trainable parameters, it excels at image recognition and is an effective alternative to building a model from scratch.Compared to other pre-trained models, ResNet-50 has excellent generalization performance and fewer errors on recognition tasks, making it a valuable tool.

EfficientNet. The EfficientNet model employs a technique
called the compound coefficient to scale models efficiently.Instead of randomly scaling up width, depth, or resolution, the compound scaling method scales all dimensions uniformly using a fixed set of scaling coefficients.By using this method and AutoML, the authors of EfficientNet created seven models of different dimensions, which achieved better accuracy than most convolutional neural networks while also being much more efficient.

Evaluation Indexes
An extensive assessment of deep learning algorithms is conducted, and a range of evaluation metrics is employed.These metrics encompass precision, recall, F1-score, accuracy, model size, parameters, training time, and testing time.This comprehensive approach allows for a thorough evaluation of the algorithms' performance: where, TP, TN, FP, and FN represent the counts of true positive, true negative, false-positive, and false negative samples, respectively.Precision quantifies the proportion of predicted positive samples that are actually positive.Recall evaluates how accurately all positive samples can be predicted as positive.The F1-score combines precision and recall into a single measure.Accuracy provides an overall assessment of sample prediction quality.Additionally, model size, parameters, training time, and testing time are commonly utilized to gauge the performance of the best model.

EXPERIMENTAL RESULT AND ANALYSIS
In this study, the application of data augmentation and the utilization of deep learning algorithms are executed within the deep learning framework provided by Keras, with Python as the primary programming language.The experimental hardware configuration specifically involves the deployment of a Tesla V100-SXM2-16GB graphics card, and the computing resources were harnessed through Google Colab.The comparison of model training results, as depicted in Figure 2.

Performance comparison between classification models
From Table 3, the model selection process culminates in the discernment of YoloV8n-cls as the optimal choice, underpinned by a multifaceted rationale.Of paramount significance is its superlative efficiency and expeditiousness, notable for its markedly petite model proportions (5.5 MB) and a parsimonious parameter count (1.44 million), juxtaposed with its peers.These attributes afford heightened resource efficiency throughout the model's lifecycle, encompassing training and inference, thereby yielding the swiftest runtimes (170.67 seconds for training and 9.68 seconds for testing).Additionally, YoloV8n-cls consistently manifests robust classification prowess, manifested in its commendable accuracy of 98.68% and a commendable F1-score of 0.98, substantiating its mettle in conferring precise and dependable predictions.This amalgamation of efficiency, performance, and diminutive deployment dimensions firmly situates YoloV8n-cls as the sine qua non for applications necessitating a judicious equilibrium between accuracy and resource optimization, particularly within the purview of real-time or edge computing domains.
In addition, we have also conducted experiments on YOLOv8 pretrained classify models, as shown in Figure 3.The training result   Consequently, following this criterion, YoloV8s emerges as the preeminent model among the alternatives presented, based on its performance as indicated by validation loss.It is essential to underscore that while validation loss serves as a pivotal metric, the ultimate selection of the most suitable model may also hinge on other variables, including the particular objectives of the task, the availability of computational resources, and the balance between model intricacy and performance optimization.

DISCUSSION AND CONCLUSIONS
The effective management of rubber leaf disease presents a significant challenge to the rubber plantation industry, necessitating the adoption of advanced detection and classification methods.The previous section provides a comprehensive overview of the evolving landscape of disease detection techniques, with a particular focus on deep learning approaches in plant pathology.In recent years, convolutional neural networks (CNNs) have emerged as a prominent machine learning approach in the field of plant disease detection [26,27].These CNNs have demonstrated their efficacy in handling diverse image types, including those captured through visible light cameras and sensors capturing non-visible wavelengths [28].Researchers have explored a variety of classification methodologies, including supervised, unsupervised, and semi-supervised techniques, employing methods such as transfer learning, support vector machines, k-nearest neighbor, and random forest algorithms in plant pathology [29].The integration of computer vision and CNNs, especially emphasizing the use of deep learning models, holds significant potential for advancing disease detection in plants [30].
However, it's important to note that while these techniques show promise, they are not without their limitations, including issues related to accuracy, speed, and cost-effectiveness.Several recent studies have demonstrated the capabilities of deep learning in disease detection and classification.For instance, Chenghai et al. introduced the DISE-Net model for maize small leaf spot disease classification, achieving remarkable accuracy compared to classical models [31].Similarly, Gnanavel et al. employed deep learning techniques to detect and categorize tomato leaf diseases, showcasing high training and testing accuracy [32].These studies underline the potential of deep learning in disease detection and its ability to support timely disease treatment, leading to improvements in crop quality and quantity.Furthermore, machine vision technology and CNNs have been applied to detect and identify major rice diseases, as demonstrated by Kantip et al. [33].These techniques have shown their impact on enhancing agricultural production and sustaining global food security.
In the context of rubber leaf disease, recent studies have explored deep learning-based approaches.For instance, Tiwei et al. introduced the GMA-Net model for rubber leaf disease recognition, achieving remarkable recognition accuracies [34].Additionally, other researchers have evaluated various CNN models in diagnosing and identifying rubber leaf diseases in Thailand, highlighting the feasibility and potential of CNNs in detecting and classifying rubber leaf diseases at various stages and severities.These advancements pave the way for developing diagnostic systems and mobile applications to assist farmers in disease management.In conclusion, the studies reviewed in this discussion underscore the effectiveness of deep learning techniques, particularly CNNs, in the early detection and accurate classification of plant diseases, including rubber leaf disease.The ongoing progress in deep learning holds the promise of transforming disease management practices in agriculture, enhancing crop yield, and contributing to sustainable agricultural practices.Future research endeavors are expected to focus on model optimization, mobile-based applications, and the expansion of datasets to encompass diverse types of rubber leaf diseases and other crop ailments.
In this study, we have explored the evolving landscape of disease detection and classification methodologies in plant pathology, with a specific emphasis on rubber leaf disease detection.The adoption of advanced techniques, including convolutional neural networks (CNNs), has shown significant promise in enhancing disease management practices in agriculture.Our experiments have culminated in the selection of the YoloV8n-cls model as the optimal choice for rubber leaf disease classification.This decision is supported by its notable efficiency, including a compact model size and modest parameter count, resulting in resource-efficient training and inference phases.Furthermore, the YoloV8n-cls model consistently demonstrates robust classification performance, with commendable accuracy and F1-score.The findings of our study contribute to the growing body of evidence showcasing the potential of deep learning techniques, particularly CNNs, in disease detection and classification.These advancements hold the promise of revolutionizing disease management practices in agriculture, ultimately leading to improvements in crop yield and contributing to sustainable agricultural practices.The constraints of this research lie in the seasonal and temporal occurrence of diseases in rubber.Over the course of this study, a dataset comprising 734 images was acquired from pertinent government agencies, categorized into three classes: healthy, powdery mildew, and new disease.Nevertheless, the outcomes and efficiency of rubber leaf disease screening are deemed satisfactory and practical for implementation.
As the field of deep learning in plant pathology continues to progress, future research endeavors will focus on further model optimization, the development of user-friendly mobile applications for farmers, and the expansion of datasets to encompass a wide range of rubber leaf diseases and other agricultural ailments.This collective effort will contribute to the continued advancement of agricultural practices and global food security.In the next stages, we plan to utilize the findings from this research to create a chat system integrated into the Line Platform.This system will serve as a diagnostic tool for the early classification of rubber leaf diseases, benefiting not only rubber plantation farmers but also academic researchers and relevant authorities.also express our deep gratitude to the Rubber Authority of Thailand and the Rubber Research Center for providing the essential rubber leaf image dataset, without which our research would not have been feasible.Additionally, we are also deeply thankful to Dr. Akadej Udomchaiporn, Head of the KMITL Digital Analytics and Intelligence Center, Faculty of Science at King Mongkut's Institute of Technology Ladkrabang and Assoc.Prof. Dr. Narit Thaochan from the Faculty of Natural Resources at Prince of Songkla University for their unwavering technical support and expertise throughout the course of this study.Their guidance and insights have been instrumental in shaping our research efforts.

Table 1 :
Photographs of the different 3 types of rubber leaf Healthy Powdery mildew New disease to classical models such as VGG16, ResNet50, InceptionV3, Mo-bileNetv1, MobileNetv2, and DenseNet121 [31].Similarly, Gnanavel Sakkarvarthi et al. employed deep learning techniques to detect and categorize tomato leaf diseases, achieving a training accuracy of 98% and a testing accuracy of 88.17%

Figure 2 :
Figure 2: comparison between classification models

Table 2 :
Typical symptoms of rubber leafHealthyVibrant and disease-free rubber leaves are characterized by a rich green hue, showcasing a smooth surface devoid of any pathological markings, with well-defined veins readily discernible Powdery Mildew Disease At the onset, these lesions manifest as diminutive, dispersed silver-white spots, characterized by a web-like arrangement of hyphae, distributed across the leaf's upper or lower surface.

Table 3 :
Comparison of classification models